Beruflich Dokumente
Kultur Dokumente
ABSTRACT
Neuro-evolution and computational neuroscience are two sci-
entific domains that produce surprisingly different artificial
neural networks. Inspired by the “toolbox” used by neu-
roscientists to create their models, this paper argues two
main points: (1) neural maps (spatially-organized identi-
cal neurons) should be the building blocks to evolve neural
networks able to perform cognitive functions and (2) well-
identified modules of the brain for which there exists com-
putational neuroscience models provide well-defined bench-
marks for neuro-evolution.
To support these claims, a method to evolve networks
of neural maps is introduced then applied to evolve neu-
ral networks with a similar functionality to basal ganglia in
animals (i.e. action selection). Results show that: (1) the
map-based encoding easily achieves this task while a direct
encoding never solves it; (2) this encoding is independent of
the size of maps and can therefore be used to evolve large
and brain-like neural networks; (3) the failure of direct en-
coding to solve the task validates the relevance of action
selection as a benchmark for neuro-evolution.
ues). Three structural mutation operators, inspired by those Similarly, connections are labeled with four evolved real
defined in NEAT [44], have been designed: numbers {c, k, Γ, σ} ∈ [0, 1)4 that are interpreted as follows:
• addition of a node on an existent connection, with ran-
1 to 1, if c < 0.5
dom labels; the connection is split in two and the two Scheme: (10)
1 to all, otherwise
parts keep the same labels;
Gaussian if c > 0.5 and k < 0.5
• removal of a random node and its associated connec- Function: (11)
constant otherwise
tions;
Γ if c > 0.5 and k < 0.5
• addition/removal of a connection between two random Parameter 1: 1 (12)
5
w (synaptic weight) otherwise.
neurons.
σ if c > 0.5 and k < 0.5
Nodes and connections can be labeled by a list of real pa- Parameter 2: (13)
otherwise: ignored
rameters that represent weights, threshold, neuron type, ...
These parameters are mutated using polynomial mutation [7]. 3.3 Classic Direct Encoding
To initialize a random graph, a random number of nodes
The previously described evolvable labeled graph can also
(with random labels) are first created then connected by a
be employed in a more classic fashion to directly define a
random number of connections (also with random labels).
neural network. This leads to a direct encoding of neural
Connections cannot be doubled. Last, graphs are simplified
networks. In this case, each node describes a neuron (instead
by removing each sub-graph not connected to both an input
of a map). Labels used to describe neurons are the same as
and an output.
those used to describe maps (time constant, etc.). Connec-
3.2.2 Map-based Encoding tions are labeled with a single real number interpreted as
the synaptic weight. This direct encoding will be used as a
To evolve a network of maps inspired by the neuroscience
reference encoding in the next section.
toolbox, the previously defined operators are employed with
a particular labeling of nodes and connections (figure 3).
This representation is independent of the size of maps, that 4. EXPERIMENTS
is the same genotype can be developed into a neural network
with an arbitrary number of neurons. 4.1 Action Selection in Basal Ganglia
In the current implementation, nodes are labeled with four Having introduced the map-based encoding, we now focus
evolved numbers {e, v, T, a} ∈ [0, 1)4 that are scaled or bi- on evaluating its ability to generate brain-like structures; to
narized to obtain nodes’ parameters. They are interpreted that aim, we chose to reproduce the function performed by
as follows3 : the basal ganglia (BG).
More precisely, the basal ganglia are a set of intercon-
inhibitory if e < 0.5 nected subcortical nuclei [36], that are thought to be in-
Type: (6)
excitatory otherwise. volved in action selection [37, 34], i.e. the problem, for an
Time constant (τ ): τ = vbτ 0 ×4c (7) agent, of choosing which action to perform within a reper-
−3 −3 −3
toire, given internal and external sensory information, in
where v = [5 · 10 , 10 , 20 · 10 , order to achieve its goals. Solving this generic resource al-
40 · 10−3 ] location problem seems to be a central cognitive function as
Threshold: 10 · (T − 0.5) (8) the BG circuitry is common to all vertebrates, and as it con-
tains duplicated circuits [1] dedicated to many critical func-
isolated neuron if a > 0.75 tions (skeleton movements, eye movements, sequence learn-
Map type: (9)
standard map otherwise. ing, working memory, navigation, planning, etc.). The BG
3
bxc denotes the floor of x. performs the two main aspects of action selection: the cen-
tral process of selection of one action only among conflicting 4.2.1 Objective Function
ones, similar to a winner-takes-all (WTA), and the learn- The main objective function (fitness) aims at checking
ing process necessary to bias the selection process towards that the maximum salience corresponds to the minimum
the most profitable option. We focus here on the WTA-like activation in the corresponding output neuron (and conse-
process. quently to the less inhibited action). Furthermore, we are
The BG circuits are subdivided in parallel channels, which interested in the best contrast possible, that is the selected
are supposed to be associated to the actions of the reper- action should be as little inhibited as possible and the other
toire, and which receive convergent inputs from the cortex. actions should be as inhibited as possible. To formulate this
At rest, the BG output is tonically active and inhibits its tar- idea, let first define min (v) (the index of the maximum value
gets. The selection of an action, caused by a strong conver- of v) and mout (x, v) (the index of the minimum value of the
gent input on the associated channel signaling its urgency, output vector, for the input v):
causes a pause of the channel output inhibition, allowing the
activation of the targeted brain region [5]. Various models of
this WTA process have been proposed; we consider here the min (v) = arg max(vi ) (18)
i∈1,··· ,k
CBG model proposed in [15], which derives from the new
functional interpretation of the circuitry proposed by [17], mout (x, v) = arg min(γc (x, v)i ) (19)
i∈1,··· ,k
but adds some previously unused connections (figure 1).
In the following, we test the ability of our new map-based We also need the index of the second minimum value of the
(2)
encoding to evolve circuits exhibiting the WTA ability ex- output vector, mout (x, v):
pected from BG models, and compare the obtained results
(2)
with CBG models whose parameters4 are optimized by a mout (x, v) = arg min (x, v)(γc (x, v)i ) (20)
i∈1,··· ,k,i6=mout
basic evolutionary algorithm.
Let I a set of N random salience vectors in [0.1, 1], uni-
4.2 Fitness formly distributed. Given that an action will be selected if
In the remaining text, the following notations are used: the corresponding output channel is below γ0 (x), we define
• x: a developed individual (a neural network); the fitness to be maximized as the average selection success
(the correct channel was disinhibited) weighted by the differ-
• k: number of inputs (also the number of outputs); ence between the minimum output (the selected action) and
the second minimum output (the first inhibited action). If
• T : the maximum simulation duration;
all outputs except the minimum one are above γ0 (x), γ0 (x)
• γ(x, v, t)i : activation level of the output neuron i (i ∈ is used to compute the difference. We thus foster the maxi-
{1, ..., k}) at time t (t ∈ [0, T ]), given the input vector mum inhibition of the selected channel but let the network
v of saliences (v ∈ Rk ); free to use any value above γ0 (x) for the other channels. The
fitness is normalized such that the maximum value is 1.
• Tc (x, v): the duration of the simulation for the input
vector v (see K(x) below); 1 X
F (x) = δmin ,mout (x, v) · γc (x, v)mout ... (21)
|I| · γ0 (x) v∈I
• γc (x, v)i : activation level of the output neuron i (i ∈
(2)
{1, ..., k}) at the end of the simulation (i.e. t = Tc (x, v)); ... − min(mout (x, v), γ0 (x))
• γ0 (x) the average activation level of the output neurons where δmin ,mout (x, v) denotes the Kronecker’s delta:
when the inputs are null, at the end of the experiment:
1 if mout (x, v) = min (x, v)
k δmin ,mout (x, v) = (22)
1X 0 otherwise
γ0 (x) = γc (x, [0, · · · , 0], 1)i (14)
k i=1
In these experiments, N was fixed to N = 1000 and the
same vectors were employed to evaluate all individuals.
Each individual is simulated until its output converges to a
constant vector or until it reaches the maximum number of 4.2.2 Constraints
time-steps (t = T ). From a practical viewpoint, a neural
The search is restricted to networks that converge during
network is considered to have converged when 10 successive
the simulation time and that do not select anything if all the
outputs have a difference of less than ε (in these experiments,
input saliences are low. These restrictions are implemented
ε = 10−6 ). To compute Ts , we first define the “convergence
using three constraints.
function” K(x, t, v):
The first constraint C1 (x) ensures that γ0 (x) is greater
than a minimum value γmin , fixed to 0.09 in this work (the
0 if γ(x, v, t)i − γ(x, v, t − n)i < ε, resting activity of the CBG model is 0.096). This means
K(x, t, v) = ∀n ∈ {1, · · · , 10}, ∀i ∈ {1, · · · , k} that all the actions should be sufficiently inhibited if all the
1 otherwise saliences are null.
(15)
Ts can now be defined as:
0 if γ0 (x) > γmin
C1 (x) = (23)
Ts (x, v) = t such as K(x, t, v) = 1, t ∈ [0, · · · , T ](16) 1 otherwise
Tc = min(T, Ts (x, v)) (17) The second constraint C2 (x) ensures that small values are
4 filtered (this behavior is observed in animals’ basal ganglia)
the CBG is parameterized by 22 synaptic weights and 3
neuron thresholds by sampling random saliences in [0, 0.1] and checking that
Table 1: Summary of launched experiments.
Search space Genotype Channels
1 synaptic weights of CBG 25 real numbers 6
2 topology and param. direct encoding 6
3 topology and param. map-based encoding 6
4 synaptic weights of CBG 25 real numbers 15
5 topology and param. direct encoding 15
6 topology and param. map-based encoding 15
1
0.8
the outputs are greater than γ0 (x): 0.6
0.4
0 if γc (x, v)i > γ0 (x),
0.2
C2 (x) = ∀v ∈ [0, 0.1]k , ∀i ∈ {1, · · · , k} (24)
0
1 otherwise CB ma dir CB ma dir
G p e G p e
(6) -bas ct en (15 -bas ct en
ed c. ) ed c.
Last, C3 (x) checks that the tested neural network con- (6 ( (1 (15
) 6) 5) )
verges to a constant output vector before the end of the ex-
fit > 0 mean fit.
periment, for all the tests performed on the neural network
(C1 (x), C2 (x) and F (x)):
C3 (x) = max(K(x), v) (25) Figure 5: Proportion of runs that satisfy the con-
v∈I straints (fitness > 0) after 200, 000 evaluations and
These constraints are enforced with the penalty method average fitness of those runs. All differences are sta-
[7]: an arbitrary large penalty is added to the fitness each tistically significant (Student T-test, p < 0.01) except
time a constraint is violated. Instead of maximizing F (x), between map-based (6) and map-based (15).
we thus maximize Fc (x):
Fc (x) = −K(C1 (x) + C2 (x) + C1 (x)) + F (x) (26)
10
where K is an arbitrary large constant (e.g. 10 ).
4.3 Experiments
To evaluate the relevance of the proposed approach to
solve the action selection task, we ask three questions: (1)
how well does the proposed map-based encoding (section
3.2.2) solve this task compared to an optimized CBG ([15],
section 4.1)? (2) compared to a basic direct encoding (sec-
tion 3.3)? and (3) does it scale up to high numbers of neu-
rons?
To that aim, six different experiments have been launched
(table 1), each of those consisting of 10 independent evolu-
tionary runs, each of them with a budget of 200, 000 evalu-
ations. The same evolutionary algorithm, the same fitness
and the same model of neurons were employed in all experi-
ments; only the genotype/phenotype mapping was changed.
The chosen evolutionary algorithm is a single-objective im-
plementation of NSGA-2 [7], an elitist tournament-based
evolutionary algorithm. Parameters are provided in table 2
and the source code is available online5 . Experiments were
carried in the Sferesv2 framework [32].